菜单
Multi-method consensus over spatial-domains. Fans out 5 methods in parallel, computes a SACCELERATOR-style base-clustering ranking, runs typed consensus (kmode / weighted / LCA), and emits a verified consensus report with the mandatory A-path banner per ADR 0010.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | consensus-domains |
| description | Multi-method consensus over spatial-domains. Fans out 5 methods in parallel, computes a SACCELERATOR-style base-clustering ranking, runs typed consensus (kmode / weighted / LCA), and emits a verified consensus report with the mandatory A-path banner per ADR 0010. |
| version | 0.1.0 |
| author | OmicsClaw |
| license | Apache-2.0 |
| tags | ["spatial","consensus","typed-consensus","expert-in-the-loop","saccelerator","bc-ranking","kmode","lca","weighted"] |
| requires | ["anndata","scanpy","numpy","pandas","scipy","scikit-learn","pyyaml"] |
The user has a preprocessed spatial AnnData (typically already QC'd via
spatial-preprocess) and wants a more trustworthy tissue-domain
assignment than any single method can produce — because the user knows
single-method results disagree on cancer / non-standard tissues, or
because the analysis is going to drive a downstream decision (cell-type
deconvolution, region-specific DE, paper figure).
This skill fans out spatial-domains over N method choices, computes a
typed statistical consensus, and surfaces the cross-method
disagreement explicitly. It does NOT replace spatial-domains; it
wraps it.
| Input | Format | Required |
|---|---|---|
| Preprocessed AnnData | --input <preprocessed.h5ad> (PCA + spatial graph) | yes |
| Output directory | --output <dir> | yes |
| Member list | --members banksy,graphst,sedr,leiden,spagcn | no (defaults to LLM-curated 5) |
| Run ALL eligible methods | --all | no (slower; SACCELERATOR-style benchmark mode) |
| Target cluster count | --n-clusters 7 | no (defaults to median across members) |
| Pre-run plan confirmation | --confirm-plan | no (default off) |
| Non-interactive BC picker | --non-interactive | no (forces top-K by score) |
| Score weights | --alpha 0.6 --beta 0.4 | no (ADR 0011 defaults) |
| Class-imbalance cap | --max-class-frac 0.8 | no |
| LLM judge veto/reweight | --llm-judge | no (default deterministic) |
| Operator | --operator {kmode,weighted,lca} | no (default kmode) |
| Seed | --seed 0 | no |
| Per-member timeout (s) | --timeout 600 | no |
| Concurrency cap | --max-parallel 4 | no |
| Output | Path | Notes |
|---|---|---|
| Verified consensus labels | consensus_labels.tsv | columns observation,consensus_<operator> |
| Per-member labels (raw) | member_<name>/figure_data/spatial_*.csv | passed through from spatial-domains |
| Cross-method NMI matrix | cross_method_nmi.csv | square matrix per member |
| Composite member scores | member_scores.csv | ADR 0011 schema |
| Markdown report | report.md | starts with [A: Verified consensus] (non-configurable) |
| Plan + audit trail | plan.json | LLM rationale + chosen operator + filtered members |
runtime/consensus/plan.propose_members reads
skills/spatial/spatial-domains/parameters.yaml param_hints,
queries the evaluation-chair LLM (or falls back deterministically),
produces N PlannedMember entries.runtime/consensus/team.run_team invokes
omicsclaw.skill.runner.run_skill("spatial-domains", ...) per
member with max_parallel = min(N, cpu_count//2, 4) and a 600 s
per-member timeout. cancel_event is propagated through.runtime/consensus/scoring.score_all_members ranks
survivors by composite alpha * cross_NMI + beta * mean_local_purity
with the max_class_frac > 0.8 hard filter.--non-interactive), accept the default.kmode / weighted / lca) on the selected base clusterings.report.md starting with the mandatory ADR 0010
banner; persist plan.json for audit; ready for graph-memory
storage under analysis://typed/<run_id>.InsufficientSurvivorsError and does
NOT silently downgrade to narrative consensus. Re-run with
--members adjusted or fall back to the dedicated narrative skill
(when shipped).[A: Verified consensus] header
is enforced by runtime/consensus/dispatch.output_banner. Do not
edit report.md to strip it before distribution.--n-clusters defaults to the median across members, not 7.
Override only when you have prior k from histology / known anatomy.--operator kmode to bypass.requires_preprocessed: true — the underlying spatial-domains
members expect obsm["X_pca"] and obsm["spatial"] populated. Run
spatial-preprocess first.# Minimal interactive run (CLI surface) — LLM picks 5, you confirm BCs
oc run consensus-domains --input preprocessed.h5ad --output out/
# Non-interactive (server / scripted)
oc run consensus-domains --input preprocessed.h5ad --output out/ \
--non-interactive
# Explicit members + weighted operator
oc run consensus-domains --input preprocessed.h5ad --output out/ \
--members banksy,graphst,sedr,leiden,spagcn \
--operator weighted --n-clusters 7
# SACCELERATOR-style benchmark (run ALL eligible methods)
oc run consensus-domains --input preprocessed.h5ad --output out/ --all
omicsclaw/runtime/consensus/ — runtime moduleexamples/consensus_benchmark/ — DLPFC 151673 hero benchmarkLLM-grounded biological interpretation of a verified typed consensus run. Reads the typed run dir + the original adata, runs inline per-cluster DE, looks up markers in a bundled tissue-keyed marker DB, and asks the chair LLM to (γ) name each cluster's likely cell type with mandatory marker citations and (β) recommend top-3 next-step skills with mandatory evidence_refs. Output banner [A+I: Interpreted on verified consensus]. Failure-mode contract per ADR 0012.
Multi-resolution typed consensus over sc-clustering. Fans out leiden / louvain at several resolutions in parallel, scores members by silhouette + cross-method NMI, runs kmode / weighted / LCA consensus on the surviving base clusterings, and emits a verified report carrying the mandatory A-path banner per ADR 0010.
Load when scaffolding a NEW OmicsClaw skill from a natural-language request — generates the skill directory layout (SKILL.md, parameters.yaml, references/, tests/) under the chosen domain. Skip when modifying an existing skill (edit its files directly) or when only routing a query (use `orchestrator`).
Load when routing a natural-language omics query to the correct domain skill across spatial / singlecell / genomics / proteomics / metabolomics / bulkrna domains via keyword / LLM / hybrid matching. Skip when the target skill is already known — invoke that skill directly.
Load when copying this directory to bootstrap a new OmicsClaw v2 skill (rename, fill in, then `git add`). Skip when an existing skill already covers the request.
Load when extracting gene programs (NMF / cNMF factorisation) and per-cell program usage scores from a non-negative scRNA AnnData. Skip when ranking marker genes per cluster (use sc-markers) or for inferring TF → target regulons (use sc-grn).